Visual features underlying neural selectivity for natural stimuli

Neural activity is selective for the identity and semantic category of natural visual stimuli. In human neuroimaging, this selectivity can be investigated using multivariate pattern analysis. The resulting neural information is thought to reflect neural processes underlying object recognition. However, due to the complexity of natural stimuli, it remains unclear to what extent different visual features drive this information. To address this, we presented natural stimuli from different categories as well as several degraded versions, in which specific visual features were selectively manipulated, to human participants while recording magnetoencephalography (MEG). We applied multivariate pattern analysis and decomposed the resulting high-dimensional neural representational geometry into its distinct contributions from retinotopy, total spatial frequency and orientation energy, global shape, local shape, and textural correlation structure. These features contributed differentially to the neural information, with retinotopy and global shape being the strongest features underlying image and category information, respectively. Our results demonstrate how multiple statistical features of visual stimuli jointly and dynamically account for neural image and category selectivity, providing a basis for a better understanding of human object recognition.